Add RDNA3/RDNA4 tensor core support

Add MMA (Matrix Multiply-Accumulate) tensor core support for AMD RDNA3
and RDNA4 GPUs by splitting AMD GPU handling into separate RDNA and CDNA
code paths with architecture-appropriate shapes.

RDNA1 and RDNA2 are explicitly blocked with compile-time constraints as
they have limited tensor core capabilities and require fallback
implementations not yet implemented.

This enables RDNA3 GPUs (RX 7000 series, W7900) and RDNA4 GPUs to use
their tensor cores for matrix operations.

Author: mcgrof

max/kernels/src/layout/tensor_core.mojo

@@ -54,6 +54,16 @@ from sys import (
     size_of,
 )
 
+from sys.info import (
+    _is_amd_rdna,
+    _is_amd_rdna1,
+    _is_amd_rdna2,
+    _is_amd_rdna3,
+    _is_amd_rdna4,
+    _is_amd_cdna,
+)
+
+
 from gpu import WARP_SIZE, lane_id, thread_idx
 from gpu.intrinsics import lop
 from gpu.memory import AddressSpace
@@ -1391,15 +1401,41 @@ fn get_mma_shape[
     else:
 
         @parameter
-        if accum_type is DType.float32 and input_type is DType.float32:
-            return shape_16x16x4
-        elif accum_type is DType.float32 and input_type.is_half_float():
-            return shape_16x16x16
-        elif accum_type is DType.float32 and input_type.is_float8():
-            return shape_16x16x32
+        if _is_amd_rdna():
+
+            @parameter
+            if _is_amd_rdna1() or _is_amd_rdna2():
+                constrained[
+                    False,
+                    (
+                        "RDNA1/RDNA2 tensor core support requires fallback"
+                        " paths (not yet implemented)"
+                    ),
+                ]()
+                return shape_null
+
+            @parameter
+            if accum_type is DType.float32 and input_type is DType.float32:
+                return shape_16x16x16
+            elif accum_type is DType.float32 and input_type.is_half_float():
+                return shape_16x16x16
+            elif accum_type is DType.float32 and input_type.is_float8():
+                return shape_16x16x32
+            else:
+                constrained[False, "Unsupported RDNA mma shape."]()
+                return shape_null
         else:
-            constrained[False, "Unsupported mma shape."]()
-            return shape_null
+
+            @parameter
+            if accum_type is DType.float32 and input_type is DType.float32:
+                return shape_16x16x4
+            elif accum_type is DType.float32 and input_type.is_half_float():
+                return shape_16x16x16
+            elif accum_type is DType.float32 and input_type.is_float8():
+                return shape_16x16x32
+            else:
+                constrained[False, "Unsupported CDNA mma shape."]()
+                return shape_null
 
 
 @always_inline